TraceSpec

The TraceSpec team built the TraceSpec-Traj dataset, containing 10,000 training states and 2,500 test states, covering two task types: browsing and software engineering.

Analysis of five speculative decoding baselines found that the aggregate acceptance rate may rise with context length, but in matched late browsing actions, the acceptance rate is actually declining.

The core method of the TraceSpec framework combines dynamic execution with fixed-prefix replay, trains on intermediate assistant turns, and uses state-conditioned prefix-aware acceptance gap estimation to allocate prediction budgets.

Fixed-prefix replay is an evaluation technique: given a prefix in the agent's interaction history and holding it fixed, the draft model generates subsequent tokens multiple times under identical conditions, and the target model's acceptance is tallied to eliminate randomness noise.

State-conditioned prefix-aware acceptance gap estimation refers to estimating the acceptance gap between the draft model and the target model based on the current trajectory state (such as task type, turn, and context length), and dynamically allocating prediction budgets accordingly, skewing toward states with large acceptance gaps and high supervision demand.

Across three target model sizes and two temperature settings, TraceSpec improves replay throughput by 12.5% to 22.6% compared with its DSpark baseline.

Ablation experiments on Qwen3-4B show that adaptive allocation improves acceptance length by 3.3% to 10.0% compared with fixed full-turn supervision.

TraceSpec's ablation experiments confirm the role of state conditioning and consecutive-token agreement in improving draft model coherence.

TraceSpec provides a set of evaluation resources and trajectory findings that connect agent execution analysis with practical draft improvement, and trajectory-conditioned speculative decoding optimization is expected to reduce inference latency.